Fig 1.
The heterogeneous graph of a microblog.
Fig 2.
The overall framework of the MHAN.
The upper part illustrates the overall structure of the MHAN. After feature extraction, the initial features of post nodes are processed by node-level attention, semantic-level attention, and classifier successively to acquire the final classification result Y. The procedure for feature extraction can be seen in the lower part. The initial features are extracted from posts’ appended visual and textual content.
Fig 3.
The hierarchical attention mechanism of the MHAN model.
Nodes 4–6 are neighbors of node i based on meta-path Ф1; nodes 2, 6, and 7 are neighbors of node i based on meta-path Ф2; nodes 3 and 9 are neighbors of node i based on meta-path Ф3. After the aggregation on each meta-path, embeddings with different semantics are fused to obtain the final representation of node i, which is denoted as Zi.
Table 1.
Statistics of the Weibo2016 and Weibo2021 datasets.
Table 2.
Statistics of the Rumors and Non-rumors in Weibo2016 and Weibo2021 datasets.
Table 3.
Experimental results on the Weibo2016 dataset.
Table 4.
Experimental results on the Weibo2021 dataset.
Fig 4.
The percentage of posts over the number of 3 types of meta-path-based neighbors.
The left column depicts how the proportion of posts changes in the Weibo2016 dataset, while the right column depicts that of the Weibo2021 dataset. The red line represents misclassified samples, and the blue line represents all samples.
Table 5.
Summarisation of the benefits and limitations of existing methods and MHAN.
Fig 5.
The visualization of the Weibo2016 posts and the weight coefficients of the meta-path PUP.
The output node features obtained by the MHAN model’s attention layers are projected onto a two-dimensional plane, and the node color indicates the class of posts. The attention weight coefficients of the meta-path PUP between nodes are denoted by the edge thickness.
Fig 6.
The visualization of the Weibo2021 posts and the weight coefficients of the meta-path PUP.
The output node features obtained by the MHAN model’s attention layers are projected onto a two-dimensional plane, and the node color indicates the class of posts. The attention weight coefficients of the meta-path PUP between nodes are denoted by the edge thickness.